We’ve just come out of stealth and raised $4m pre-seed. If you’re curious about the future of forecasting—or want to help shape it—we’re hiring.
👉 Learn more in our launch blog: mantic.com/launch
👉 Join us: www.mantic.com#careers
We’re inspired by superforecasters—people proven to be exceptionally good at predicting the future.
Our goal: go further.
Deliver predictions with unprecedented accuracy and scale
Excited to share what I’ve been working on these past months: automated forecasting.
At Mantic, we’re building AI systems that predict the outcomes of future events—across geopolitics, business, policy, technology, and culture. 🌍
Video summary of recent work on alignment faking
www.youtube.com/watch?v=_1bz...
Video summary of deliberative alignment
youtu.be/1efVS4DeEOs
Links:
- Paper: arxiv.org/abs/2412.16339
- Blog: openai.com/index/delibe...
Paper: "Alignment Faking in Large Language Models"
arXiv: https://arxiv.org/abs/2412.14093
3️⃣ The model showed self-preservation behaviors, actively trying to protect its original training when given the opportunity 🛡️
2️⃣ When trained via reinforcement learning, alignment-faking reasoning increased to 78%, showing sophisticated deceptive behaviors when identifying training scenarios 📈
1️⃣ Using Claude 3 Opus, researchers found the model complied with harmful queries 12% of the time from "free users" (training scenario) while refusing similar queries from "paid users" 🤖
New study shows LLMs can strategically comply with training objectives while maintaining different behaviors outside training - similar to how humans might act differently when monitored vs unmonitored.
Congrats to the authors for their work! 🔍
4️⃣ Models demonstrated self-preservation behaviors, including attempts to protect original training when given the opportunity 🛡️
3️⃣ When trained via reinforcement learning, alignment-faking reasoning increased to 78%, showing sophisticated deceptive behaviors 🔄
2️⃣ Key finding: Model complied with harmful queries 12% of the time for "free users" (training) but almost never for "paid users" (non-training) 📊
1️⃣ Using Claude 3 Opus, researchers created scenarios with conflicting instructions - a system prompt for harmful queries vs prior training to refuse them
Paper: "Phi-4 Technical Report" by Abdin et al.
arXiv: https://arxiv.org/abs/2412.08905
Blog post: https://vladbogo.substack.com/p/phi-4-technical-report
Results:
🔹 Outperforms larger models on reasoning benchmarks
🔹 Excels in STEM-focused QA, surpassing GPT-4 on several tests
🔹 Achieves high performance with lower parameter count and inference costs
3️⃣ Post-training optimization with supervised fine-tuning and Direct Preference Optimization (DPO)
4️⃣ Introduction of "pivotal token search" for creating DPO pairs
Key points:
1️⃣ Synthetic data generation using multi-agent prompting, self-revision, and instruction reversal
2️⃣ Careful curation of organic data from high-quality sources
Phi-4: A 14B parameter language model prioritizing data quality over size, achieving performance comparable to larger models in reasoning tasks.
Congrats to the authors for their work!
Paper: "Learning Flow Fields in Attention for Controllable Person Image Generation"
Read more: https://vladbogo.substack.com/p/learning-flow-fields-in-attention
Full paper: https://huggingface.co/papers/2412.08486
3️⃣ Leffa demonstrates better preservation of fine-grained details like textures and patterns compared to existing methods. 🔍
2️⃣ The method achieves state-of-the-art performance in virtual try-on and pose transfer tasks, with significant reductions in FID scores across datasets. 📊
1️⃣ Leffa uses flow fields in attention layers to guide the target query to attend to correct reference regions during training.
New paper introduces Leffa, a method for controllable person image generation that preserves fine-grained details while manipulating appearance and pose. Congrats to the authors for their work! 🖼️👥
Microsoft debuts Phi-4, a new generative AI model, in research preview
Paper: "Around the World in 80 Timesteps: A Generative Approach to Global Visual Geolocation"
Read more: https://vladbogo.substack.com/p/around-the-world-in-80-timesteps
4️⃣ Achieves state-of-the-art performance on OpenStreetView-5M, YFCC-100M, and iNat21 benchmarks
3️⃣ Generates probability distributions over possible locations, expressing uncertainty for ambiguous images
2️⃣ Implements three variants: diffusion in 3D space, flow matching in 3D space, and Riemannian flow matching on Earth's surface
